Obfuscated Text Detection in Digital Ads via Neural OCR
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for identifying malicious content in digital advertisements are inefficient and can be circumvented by advertisers using characters from different Unicode blocks, leading to cumbersome human review and inaccurate automated detection.
Innovation Solution
An online system uses a neural network-based obfuscation detection model trained with a text obfuscation library to identify and prevent the presentation of obfuscated text in advertisements by performing optical character recognition and calculating the probability of obfuscation in character embeddings.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If conventional automated systems analyze advertisement content to identify malicious text, then detection speed is improved, but detection accuracy deteriorates due to obfuscated text using different Unicode blocks
Solution Approach 1:
The patent introduces an intermediary processing layer that converts obfuscated text from different Unicode blocks into a standardized representation before analysis. The system renders the advertisement content to a visual format, performs optical character recognition to extract text, and then analyzes character embeddings to detect obfuscation patterns. This intermediary rendering and OCR process acts as a mediator that translates various Unicode representations into a common format that the detection model can accurately analyze, thereby maintaining both high detection speed and high detection accuracy.
2Measurement precision
If manual review methods are used to identify obfuscated text, then detection accuracy is improved, but processing time increases significantly
Solution Approach 1:
The patent replaces the mechanical manual review system with an automated neural network-based detection system. The system uses a trained model that analyzes character embeddings and calculates obfuscation probability automatically. This substitution eliminates the need for human reviewers while maintaining high detection accuracy, as the neural network has been trained on diverse obfuscation patterns. The automated system processes advertisements at machine speed, dramatically reducing processing time compared to manual review while preserving or even improving detection accuracy through consistent application of the detection algorithm.
3Adaptability or versatility
If advertisers use characters from different Unicode blocks to obfuscate text, then evasion of detection is improved, but system complexity increases
Solution Approach 1:
The patent addresses Unicode-based obfuscation by changing the parameter space in which text is analyzed. Instead of analyzing raw Unicode code points directly, the system renders text to visual characters and extracts features based on visual appearance and character embeddings. This parameter transformation allows the system to detect obfuscation regardless of which Unicode block the original characters came from. The neural network model learns to recognize obfuscation patterns in the transformed parameter space, effectively neutralizing the evasion capability provided by different Unicode blocks without requiring the system to handle each Unicode block separately, thus managing complexity.
Data Source
AI summary
An online system receives digital content and determines whether the digital content includes malicious content, such as obfuscated text, before presenting the digital content to a user. To determine whether the digital content contains malicious content, the online system renders the digital content. The online system performs optical character recognition on the content. The online system uses an obfuscation machine learning model to identify obfuscated text. The online system may deobfuscate the obfuscated text. The online system may prevent presentation of the digital content in response to detecting obfuscated text.